Silicon Graphics Origin 350, Руководство пользователя

Страница: 94 / 274

68 007-4566-001
2: System Overview
Functional Architecture
The Origin 350 server system is based on the SGI NUMAflex architecture, which is the
third-generation shared-memory system architecture that is the basis of SGI HPC servers
and supercomputers. The NUMAflex architecture is specifically engineered to provide
technical professionals with superior performance and scalability in a design that is easy
to deploy, program, and manage. It has the following features:
Shared access of processors, memory, and I/O. The Bedrock ASICs and the NUMAlink
interconnect functions of the NUMAflex architecture enable applications to share
processors, memory, and I/O devices.
• Each Bedrock ASIC in the system is an 8-input by 6-output crossbar that acts as the
memory controller between processors and memory in the system for both local
and remote memory accesses.
• The NUMAlink interconnect channels information between all the modules in the
system to create a single contiguous memory in the system of up to 64 GB and
enables every processor in a system direct access to every I/O slot in the system.
Together, the Bedrock ASICs and the NUMAlink interconnect enable efficient access to
processors, local and remote memory, and I/O devices without the bottlenecks
associated with switches, backplanes, and other commodity interconnect technologies.
System scalability. The NUMAflex architecture incorporates a low-latency,
high-bandwidth interconnect that is designed to maintain performance as you scale
system computing, I/O, and storage functions. For example, the computing dimension
in some system configurations can range from 4 to 32 processors in a single system image
(SSI). To increase the number of processors beyond 32, you can cluster Origin 350 server
systems together. The maximum number of processors in a clustered configuration is
128.
Efficient resource management. The NUMAflex architecture is designed to run complex
models and, because the entire memory space is shared, large models can fit into
memory with no programming restrictions. Rather than waiting for all of the processors
to complete their assigned tasks, the system dynamically reallocates memory, resulting
in faster time to solution.